We present a method that, for the first time in a broad coverage setting, uses natural language generation to automatically construct disambiguating paraphrases for structurally ambiguous sentences. By simply asking naive annotators to clarify which paraphrase is closer in meaning to the original sentence, the resulting paraphrases can potentially enable meaning judgments for parser training and domain adaptation to be crowd-sourced on a massive scale. To validate the method, we demonstrate that meaning judgments crowd-sourced in this way via Amazon Mechanical Turk have reasonably high accuracy-e.g. 80%, given a strong majority choice between two paraphrases-with accuracy increasing as the level of agreement among annotators increases. We also show that even with just the limited validation data gathered to date, the crowd-sourced judgments make it possible to retrain a parser to achieve significantly higher accuracy in a novel domain. We conclude with lessons learned for gathering such judgments on a much larger scale.

We investigate whether parsers can be used for self-monitoring in surface realization in order to avoid egregious errors involving "vicious" ambiguities, namely those where the intended interpretation fails to be considerably more likely than alternative ones. Using parse accuracy in a simple reranking strategy for selfmonitoring, we find that with a stateof-the-art averaged perceptron realization ranking model, BLEU scores cannot be improved with any of the well-known Treebank parsers we tested, since these parsers too often make errors that human readers would be unlikely to make. However, by using an SVM ranker to combine the realizer's model score together with features from multiple parsers, including ones designed to make the ranker more robust to parsing mistakes, we show that significant increases in BLEU scores can be achieved. Moreover, via a targeted manual analysis, we demonstrate that the SVM reranker frequently manages to avoid vicious ambiguities, while its ranking errors tend to affect fluency much more often than adequacy.

Categorial grammars are attractive because they have a clear account of unbounded dependencies. This accounting is especially important in Mandarin Chinese which makes extensive usage of unbounded dependencies. However, parsers trained on existing categorial grammar annotations (Tse and Curran, 2010) extracted from the Penn Chinese Treebank (Xue et al., 2005) are not as accurate as those trained on the original treebank, possibly because enforcing a small set of inference rules in these grammars leads to large sets of categories, which cause sparse data problems. This work reannotates the Penn Chinese Treebank into a generalized categorial grammar which uses a larger rule set and a substantially smaller category set while retaining the capacity to model unbounded dependencies. Experimental results show a statistically significant improvement in parsing accuracy with this categorial grammar.

We present a simple, broad coverage method for clarifying the meaning of sentences with coordination ambiguities, a frequent cause of parse errors. For each of the two most likely parses involving a coordination ambiguity, we produce a disambiguating paraphrase that splits the sentence in two, with one conjunct appearing in each half, so that the span of each conjunct becomes clearer. In a validation study, we show that the method enables meaning judgments to be crowd-sourced with good reliability, achieving 83% accuracy at 80% coverage.

This paper introduces our Chinese semantic dependency parsing system for Task 9 of Se-mEval 2016. Our system has two components: a parser trained using the Berkeley Grammar Trainer on the Penn Chinese Treebank reannotated in a Generalized Categorial Grammar, and a multinomial logistic regression classifier. We first parse the data with the automatic parser to obtain predicate-argument dependencies and then we use the classifier to predict the semantic dependency labels for the predicate-argument dependency relations extracted. Although our parser is not trained directly on the task training data, our system yields the best performance for the non-local dependency recovery for the news data and comparable overall results.